GirkovArpa

This comment needs to be higher.

> You can ask them to reverse numbers or re-arrange words and they'll faceplant in the same way as soon as the input gets beyond a small threshold. Here surely there wouldn't be an issue with tokenization.

My guess is the training data contains many short pairs of forward and backward sequences, but none after a certain threshold length (due to how quickly the number of possible sequences grows with length). This would imply there's no actual reversing going on, and the LLM is instead using the training data as a lookup table.

This should work fantastic in theory, since differing vocabulary (not grammar) is the main factor that determines the difficulty of a new language. Putting off this primary obstacle so one can ease into it sounds genius to me. It also agrees with the method hyped by Steve Kaufman, where one should read and speak level-appropriate material.

Exactly. The proof they did not actually crack Bitcoin is the fact they said they did.

In the context of Condorcet's jury theorem, the percentages refer to the chance of voting for the correct outcome. Think of a legal trial and there is no ambiguity about the meaning of "50%" is.

According to Condorcet's jury theorem, a committee of 10 identical members may be smarter than a single member.

Yes. Back around 2020 they forgave me a $1000 bill for a side project I thought was running on a free tier.

I love it! Going to use this instead of Pure.CSS from now on.

> All the languages in Europe descended from the Proto-Indo-European language

Except for Basque!